<- all tokdocs

Cloud Filesystems for AI Agents: Who Already Ships One

By , founder of Mavric and Apso

View on TikTok ->

AI coding agents from Anthropic and OpenAI each work on their own copy of your files, so they cannot see each other's changes until someone pushes to Git. That part of a viral TikTok is accurate. The claim that a shared cloud filesystem for agents is missing infrastructure you have to build yourself is not: Cloudflare, AWS, Mesa and Turso had all shipped versions before the video was posted on October 8, 2026.

The video comes from Isaac Dyor (@isaacdyor), who says "We had to build this for ourselves at Compound" and describes a design built on Cloudflare Durable Objects and R2. The design is sound. The framing around it needs correcting.

Why can't AI agents see each other's files?

Each agent runs in its own environment. The video says "each agent operates on its own file system, whether that's your computer, your teammate's computer, or a cloud sandbox, which means that they can't see each other's work."

Anthropic's Claude Code documentation confirms this. Cloud sessions run "in an isolated, Anthropic-managed VM," and "The cloud VM clones your current directory's GitHub remote at your current branch, not your local checkout, so push first if you have local commits." Work comes back as a pull request. Pulling a session to your terminal gives "its own copy of the session," and shared session links show "the latest state when they open the link, but their view doesn't update in real time."

OpenAI's Codex cloud docs say the same: "Each task has its own workspace and can keep working while your computer is asleep."

Is Git a good way to share files between agents?

The video's answer is that Git works for code but is not live and "falls apart on large files like video." GitHub's own limits back this up:

  • Git warns on files larger than 50 MiB
  • Browser uploads are capped at 25 MiB
  • GitHub blocks files larger than 100 MiB
  • Anything larger needs Git Large File Storage (Git LFS)

Anthropic adds its own limits for cloud sessions. "The bundled repository must be under 100 MB," and larger repos fall back to the current branch or a squashed snapshot. Claude Code also leaves out uncommitted changes to files managed by a Git filter such as Git LFS.

Do shared cloud filesystems for agents already exist?

They do, and several shipped before the video. It opens by saying "OpenAI and Anthropic are still missing the piece of infrastructure that allows you, your teammates, and your agents to all work on the same files" and that "You basically need to build" it. Anthropic and OpenAI do not ship a live shared filesystem in their own agent products. The wider category was already shipping:

  • Cloudflare @cloudflare/computer (August 3, 2026). An agent runtime with "a durable filesystem that you can use with git repositories, storage buckets or any files you choose," with the workspace "backed by SQLite." That is the same two-layer design the video presents.
  • Cloudflare R2 FUSE mounts (November 21, 2025). "Containers now support mounting R2 buckets as FUSE (Filesystem in Userspace) volumes."
  • Cloudflare artifact-fs (Apache-2.0). A FUSE driver that mounts large Git repos and fetches file contents on demand, so "reads block only until the requested blob is fetched."
  • AWS Bedrock AgentCore Runtime (May 6, 2026). Mounts S3 and EFS into agent sessions, and "Multiple agents, or multiple sessions of the same agent, can collaborate on the same dataset."
  • Mesa (April 28, 2026). "A fully posix-compatible, durable filesystem with built-in version control," mountable with FUSE.
  • Turso agentfs (MIT). A SQLite-backed agent filesystem where "Every file operation, tool call, and state change is recorded in a SQLite database file."

How does a cloud filesystem on Cloudflare work?

The video's design splits the job in two. "A durable object keeps track of every file and every change, and R2 stores the actual bytes of the files." Each machine mounts the workspace like a local folder and streams only the parts of a file it opens.

The pieces exist as described. Cloudflare defines a Durable Object as "a special kind of Cloudflare Worker which uniquely combines compute with storage," with storage that is "strongly consistent yet fast to access." SQLite storage in Durable Objects went GA on April 7, 2025.

Can it handle hundreds of gigabytes of video?

The video says "you can work on hundreds of gigabytes of video without filling up your disk." That works only because R2 holds the bytes. R2 allows up to 5 TiB per object, with a 5 GiB single-part upload limit and multipart uploads up to 4.995 TiB across 10,000 parts.

The Durable Object side has a ceiling the video never mentions: 10 GB of SQLite per object, per Cloudflare's GA changelog. The design holds up if the Durable Object stores only metadata and version numbers, which is what the video's diagram implies.

What are the limits of mounting object storage?

The video says when anyone changes a file, "all the other computers see it instantly." Cloudflare's own docs are more cautious: "FUSE provides filesystem access to object storage. It does not provide local-disk performance or full POSIX filesystem semantics." Its example mounts the bucket read-only, and Cloudflare lists uses like loading datasets, models and shared assets.

The video does not cover what happens when two agents edit the same file at once. Per-file version numbers in its diagram are the only hint of a versioning scheme. Mesa handles concurrency differently, with "branches so agents can work in parallel without locking."

The video shows diagrams only, with no running system, code or repository. I could not find a public repo, docs page or launch post for the Compound filesystem it describes.

Bottom line

The problem is real: Claude Code and Codex give each agent its own copy, and Git struggles with large files. The claim that nobody has built the fix is wrong, since Cloudflare, AWS, Mesa and Turso shipped agent filesystems between November 2025 and August 2026. A Durable Object plus R2 design can handle large video if the Durable Object stays under its 10 GB limit. Concurrent writes are the open question the video skips.

Sources

Published October 8, 2026. Based on a TikTok by @isaacdyor. Full video notes (markdown)